Muse Spark Open ships weights, a license, and a desktop bill

Weights with a count attached

Meta's October 15, 2026 research post says Muse Spark Open is available and that the weights are released. It says the model is natively multimodal, with 816 billion parameters of which about 42 billion are active, and a one million token context. It says the model is built for coding and longer tasks. Hugging Face's same-day account adds the shape: 256 experts in each mixture-of-experts layer, 8 selected per token, 62 layers, inputs of text, image, audio, and video, and text out. The BF16 checkpoint is described there as about 1.5 terabytes. Open weights a person can fetch are not the training source.weights you can fetch.

The license is a limit, not a vibe

Meta says the weights ship under the Muse Model License with limited commercial terms. That is a license with a ceiling, which the post summarizes as free for many developers. A reader who needs to know whether a product is inside the ceiling has to read the license, not the headline that says open. This page will not collapse limited commercial terms into public domain, and it will not collapse them into a ban.the open-source desk.

Most capable is Meta's ranking

The post calls the model the most capable U.S. open-weight model to date. That is Meta's ranking of its own release. Hugging Face repeats the description. Neither page is an independent eval a second lab has published in the sources this brief uses. A benchmark table on the author's blog stays the author's table. Asked directly, How big does Meta say the model is? 816 billion parameters in total, about 42 billion active per token, and a context window of up to one million tokens. Hugging Face's account of the same release matches those three figures.an evaluation a second lab can run.

The desktop bill is two machines, or four

Meta says this is the first model in its class to run on a desktop, and then it defines the desktop: two M5 Ultra Mac Studios, 512 gigabytes each, using MLX. The hardware page also allows four machines with 256 gigabytes. Apple's MLX note describes a four-Mac cluster and says it exceeds 36 tokens per second. That speed is Apple's. The memory bill is the fact a buyer can check against a price list. One laptop is not the configuration either page describes.the hardware desk.

What the posts say was run on that desk

Meta says it tested agent-style work entirely on the machines, including code changes to an open-source Mastodon client and a video edit in DaVinci Resolve. Those are demonstrations Meta reports. They are not a public suite with a score a stranger can rerun from the blog alone. The weight files are the artifact. The demo is the company's story about the artifact. Asked directly, What speed figure is Apple's, not Meta's chart? Apple's MLX note says a four-Mac cluster running this model exceeds 36 tokens per second. That is Apple's measurement of Apple's stack. This page did not rerun it.Futureweb.

Safety language is a threshold Meta says it met

The post says the model had extra safety training because the weights are open, that it was evaluated under Meta's scaling framework, and that it met the thresholds that framework requires. Detailed results are deferred to a model card, with a fuller report described as coming later. Until that report is public, the threshold sentence is Meta's. This page will not invent a score the post did not print.

Where the weights are, and where the training set is not

Meta lists API hosts and four places to download weights. Download is the open-weight claim a reader can test by fetching a file. The post does not publish the training corpus. A weight file without the code and data that produced it is the distinction the catalog brief already draws, and this release does not erase it. Asked directly, What is the license, in Meta's words? The Muse Model License with limited commercial terms, which Meta says allows many developers to use the model for free. Open weights under that license are not the same thing as training code a reader can rebuild.

Engines are a list, not a benchmark

Meta says the model runs on SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth, with vendor kernels for several chips. A supported-engine list means a project claims a path. It does not mean each path has the same tokens per second. Apple's 36 is one path. The others are names until someone measures them. Asked directly, What desktop hardware does Meta name? Two Apple M5 Ultra Mac Studios with 512 gigabytes of unified memory each, or four with 256 gigabytes, linked so the weights fit. Meta's cookbook says macOS 26.2 or later and a direct Thunderbolt 5 link.

What this page will re-check

A license revision, a model card that changes 816 or 42, or a second lab's tokens-per-second figure on the two-Mac bill. The October 15 release stands as weights, a limited commercial license, and a hardware requirement that is two large desktops or four smaller ones. The short answer this page will stand on is this: On October 15, 2026, Meta released Muse Spark Open weights: a multimodal mixture-of-experts model it describes as 816 billion parameters with about 42 billion active, and a context of up to one million tokens. The license is Meta's Muse Model License, with limited commercial terms. Meta says the desktop path is two M5 Ultra Mac Studios with 512 gigabytes each, or four with 256.